A Saliency-based Unsupervised Method for Angiectasia Detection in Endoscopic Video Frames

The detection of angiectasia, the primary suspected lesion in patients with obscure gastrointestinal bleeding, presents a challenging problem for physicians. In this paper, we present a saliency based unsupervised method for automatic localization and detection of angiectasia in wired and capsule endoscopic images. To achieve comparable illumination in images from both modalities, image enhancement based on Retinex is performed on the capsule endoscopic images. A saliency detection algorithm has been proposed where the saliency map is formed from the processed images using two distinctness measures: pattern distinctness and color distinctness. The angiectasia specific saliency detection algorithm is able to highlight the lesion affected areas. An adaptive thresholding is performed based on the saliency peaks detected from the gradient images. The performance of the proposed method is evaluated on a dataset consisting of 3602 images, among which 968 images show the indication of angiectasia. The method achieves very high localization score (95.04%), localization precision, moderate specificity (>80%) and a very low detection latency (<0.2 s) for both imaging modalities. A comparison with state-of-the-art saliency detection methods exhibits the efficacy of proposed saliency detection algorithm for angiectasia localization and detection.